Fix sub-table of a dotted key escaping its parent header path (#557)

[earfman]

Fixes #557.

Problem

Assigning a sub-table under a dotted-key super-table nested inside a real table caused the sub-table header to lose its parent path on render:

doc = tomlkit.loads("\n[t]\na.b = 1\n")
doc["t"]["a"]["c"] = {}
tomlkit.dumps(doc)
# [t]
# a.b = 1
#
# [a.c]   <- escapes [t]; `c` is lost from t.a and reappears at the document root

parse(dumps(doc)) != doc, so this is a round-trip / data-loss bug. The same happened inside array-of-tables elements ([[servers]] produced [a.c] instead of [servers.a.c]).

Cause

In Container._render_table, prefix served two purposes at once: the relative dotted-key prefix used for leaf items (so a.b stays inline under [t]) and the absolute header path used for nested sub-tables (which need [t.a.c]). The "dotted key inside table" branch dropped the prefix entirely — right for the leaf, wrong for a real sub-table nested underneath, which then lost the t. parent path. _render_aot_table had the same issue.

Fix

Track the absolute header path separately via a new header_prefix argument. Leaf items continue to render off the relative key; nested sub-table and array-of-tables headers render off the absolute path. When a table carries a display_name (from a parsed explicit header) the two paths coincide, so existing output is unchanged.

Tests

Added regression tests covering the table case, a chained dotted parent (a.b.x — the leaf stays relative while the sub-table header is absolute), and the array-of-tables case. The full suite passes (1053 tests).

---

Found and fixed with an integrity-first pipeline; the fix was independently verified by a separate agent than the one that found it before submission.